Performance and Scaleability tracking ticket

[RudolfWeeber]

My current understanding:

• Currently it seems best to focus at the Lennard-Jones benchmark.

• Lennard-Jones (together with other short range forces) are most relevant on the particle side

• If they are faster, electrostatics will be also faster because work can be shifted away from the not-well-scaling parallel 3d FFTs

• Analysis of electrostatics and Lattice-Boltzmann might be deferred until the significant planned code changes there have come in

  • Kokkos and replaced FFT library for electrostatics
  • Switch to pull scheme and (multi)-GPU support for waLBerla backed LB

Optimizaiton targets for short-ranged forces

• Performance of the hot-loop (finding interacting particle pairs and calculating the pair forces). Improving this will also reduce communication by allowing for a larger skin in the Verlet lsit.
• Communication scheme
• Avoiding reduction(sum) of forces from particles in ghost layers by recalculating theme on all affected MPI ranks
• Particle memory layout (using struct-of-arrys)
• Load imbalances across MPI ranks
• Use MPI+X so as to avoid (most) communication on the node level

Relevant tickets

• Performance regressions since 4.0 #4866 (suspects for performance regressions since 4.0)